Resistive Switching mechanism study

Improving the Resistive Switching Property

Improved electrical endurance can be obtained in a crossbar type Pt/TiO₂/Pt structure by using a modified bias scheme. With a conventional bias scheme, for example positive set and positive reset, conical shape conduction filament having a narrower cross sectional area at the anode interface is formed in the TiO₂ memory layer. Because resistive switching occurs near the top electrode, it causes a rapid oxygen consumption leading to the poor electrical endurance. To prevent the rapid oxygen consumption, a modified bias scheme composed of a series application of positive and negative bias is used. As a conical filament shape is generated by a single bias scheme, the modified bias scheme results in an hourglass shaped conduction filament as confirmed by the HRTEM of the device after repeated switching cycles. Eventually, the oxygen consumption is effectively suppressed and the switching endurance was highly improved.

Related Papers:

K.M. Kim et al., Electrochemical and Solid-State Letters, 6, 13, G51-G53 (2010)

G. H. Kim, et al., Appl. Phys. Lett. 98, 262901 (2011)

Switching uniformity is another critical issue for resistive switching memory. The formation and rupture of conduction filament are stochastic phenomena so they give highly scattered operation parameters. The variability of the operation parameters was compared for the cases where Ru nano-dots were embedded in the vicinity of bottom or top interface region of a Pt/TiO₂/Pt resistive switching cell. Combined with ATLAS simulation results, the AFM and CAFM analysis reveal that the Ru nano-dots role as either a local field enhancer or field suppressor when inserted at the cathode or anode interface, respectively. Therefore, limiting the location where electron injection occurs at the cathode interface to a narrower region was the key factor for achieving highly improved RS uniformity.

Related Papers:

J. H. Yoon, et al., Appl. Phys. Lett. 97, 232904 (2010)

J. H. Yoon, et al., Adv. Mater. 25, 1987-1992 (2013)

Selector Employed 3D RRAM Crossbar Array Intergration

32x32 crossbar array(CBA) of >1kb RRAM was fabricated. Each cell is comprised of 1diode-1resistive switching memory stack, where a diode selector was employed for each resistive switching element to alleviate the sneak currents issue in the CBA. In order to accomplish the stack, a Pt/TiO₂/Ti/Pt diode was stacked above a Pt/TiO₂/Pt unipolar resistive switching memory. The diode selector adopted the TiO₂ deposited by ALD method with O₃ as the oxidant. The maximum rectification ratio achieved in the diode was >10⁶, while it was >10⁵ in the 1D1R device. The high rectification ratio is attributed to the high Schottky barrier at the Pt/TiO₂ and the quasi-Ohmic conduction through the TiO_2-x/Pt, which is obtained by the partial oxidation of Ti during the subsequent oxide deposition. Conduction through a local conducting paths was confirmed for the both injections, based on the electrode area dependent current density and the conducting AFM measurements.

Related Papers:

G. H. Kim et al., Adv. Funct. Mater., 11, 23, 1440  (2012)

G. H. Kim et al., Appl. Phys. Lett. 21, 100, 213508 (2012)

Double-layers stacked RRAM was demonstrated in 4x4x2 CBA. In order to avoid the thermal disruption issue during the stacking process, which is particularly critical for the diode layer embedding high chemical potential within the layer, room temperature sputtering was adopted for the deposition of every oxide layer. The room temperature fabricated Pt/TiO₂/Ti diode showed even improved property with the rectification ratio of >10⁸, which was still maintained in the 1D1R device. The enhancement in the diode rectifying property is mainly ascribed to the suppression of the partial reduction of TiO₂, which degrades the performance of Schottky barrier at the Pt/TiO₂ interface. A 2-bit operation of the 1D1R further scaled the cell size down to F²_.

Due to the low per lithographic margin of the aforementioned stacked RRAM structure, vertical RRAM structure adopting etching process as the key fabrication technique is preferred. Since the middle electrode placed in between the selector and the memory can no longer be adopted in this structure, however, a self-rectifying RRAM comprised of only the two different oxide layers is developed. The Pt/Ta₂O₅/HfO₂/Ti device showed a resistive switching based on trapping/detrapping mechanism, accompanied with high rectification ratio of ~10⁶_. Furthermore, the highly promising 3-bit operation was demonstrated.

Related Papers:

J. H. Yoon et al., Adv. Funct. Mater., 32, 24, 5086 (2014)

J. H. Yoon et al., Adv. Mater., 25, 27, 3811 (2015)

Study on Crossbar Array Architecture

The diode requirements based on the sneak currents analysis in a planar CBA has been examined. In CBA structure, the sneak current paths comprise a parallel resistor component to a selected cell. Therefore, unwanted sneak currents can interrupt the accurate reading of the selected cell. By calculating all the possible sneak current components within a CBA, the required rectification ratio and the forward current density of the diode according to memory density has been evaluated.

Related Papers :

G. H. Kim et al., Nanotechnology, 21, 385202 (2010)

Further improved from the sneak currents analysis in the planar CBA, sneak currents analysis on 3dimensional- (3D-) CBA was examined. Two ideas of realizing the 3D RRAM structure was proposed: cross-line structure, and word-plane structure. The cross-line CBA is the structure where multiple layers of planar CBA are repetitively stacked, whereas the word-plane structure is the structure where the selector and memory oxides are overlaid on a vertically erected metal sidewall structure. Particularly for the latter, the number of bit-lines sharing a word-line increases with the increasing number of layers, therefore, sneak currents problem gets severer with the increasing number of layers. Therefore, higher selector performance is required for the word-plane structure compared to that of the cross-line structure. However, the high per lithographic margin of the word-plane makes the structure highly cost-effective, which is favorable for the ultimate success of the 3D RRAM.

Related Papers :

J.Y. Seok et al., Adv. Funct. Mater., 34, 24, 5316-5339 (2014)

On the other hand, it has been recently found that the sneak currents can be problematic even during writing operation. The sneak currents may incur adverse voltage drop on the selected word- and bit- line and may also disturb the unselected cells on the sneak current paths. An analytical model for evaluating the increased writing voltage as well as the decreased writing margin due to the additional voltage drop on the selected word- and bit- line by the involvement of sneak currents has been deduced. A HSPICE simulation results provide the validity of the analytical model, while the model was applied to the 1D1R device eventually providing a plausible design strategy of the device in order for its implementation to a high density CBA RRAM.

Neuromorphic Engineering

Introduction to Neuromorphic Engineering

Neuromorphic engineering means the concept describing the use of very-large-scale integration (VLSI) systems containing electronic analogue circuits to mimic neuro-biological architectures present in the nervous system. The implementation of neuromorphic computing on the hardware level can be realized by oxide-based memristors, threshold switches and transistors.

Relaxation oscillator-based leaky integrate-and-fire

Leaky integrate-and-fire (LIF) model is the most popular model of a neuron. Leaky integrate-and-fire model consists of a capacitor like cell wall in real neurons and a resistance to make a leakage current. (Left figure) As input current is applied, voltage difference between extracellular and intracellular is increased. When the voltage reaches to threshold voltage, LIF neuron fires a spike and the voltage comes back to resting voltage.

One way to construct hardware LIF neuron is using relaxation oscillator as LIF neuron, so called relaxation oscillator-based LIF neuron (ROLIF neuron). The right figure shows structure of ROLIF neuron which uses Pearson-Anson oscillator (PAO) as relaxation oscillator. As input current is applied, voltage on threshold increased because of the capacitor. When the voltage reaches to threshold voltage, threshold switch (TS) is changed to low-resistance state (LRS) and it causes increase of V_out which means spike firing. Since TS is changed to LRS, voltage on TS becomes lower and it makes TS changed to high-resistance state again.

In ROLIF neuron, properties of threshold switch are very important. For example, the left graphs shows the properties of Pt/GeSe/Pt TS.

Top-right graph shows the spike train from ROLIF neuron. The activity, the number of spikes in a given time period(8 μs), tends to increase with V_in.

To highlight this relationship, we further identified the change in the number of spikes upon V_in (right-bottom graph). First, the graph reveals that there exists a threshold voltage for spiking (ca. 3.5 V) that is determined by V_on. At V_in below this threshold, V_TS cannot reach V_on, so no spike is evoked. Second, the number of spikes tends to gradually increase with increasing V_in in a wide Vin range. This change in the spike number is attributed to a decline in the inter-spike interval (ISI) with V_in. Third, the activity is saturated at a high V_in (>ca. 5.2 V). The capacitor’s charging and discharging is in need of non-vanishing time. Thus, this minimally required time delimits the minimum ISI, i.e., maximum spike number.

Related Paper:
H. Lim et al., Nanoscale, 18, 9629-9640 (2016)